SMGDiff: step mapping generalized diffusion model for efficient noise reduction in cardiac-gated myocardial perfusion...

Improving the image quality of cardiac gating myocardial perfusion single-photon emission computed tomography (CG MP-SPECT) is crucial for accurate diagnosis. Diffusion model (DM) has recently shown promise in MP-SPECT image denoising, but traditional DM typically require extensive computational resources and prolonged processing times. The study aims to develop and evaluate a lightweight generalized DM for efficient denoising in CG MP-SPECT images.

We propose a novel step mapping generalized diffusion model (SMGDiff) that incorporates cardiac-gated MP-SPECT images as diffusion endpoints instead of traditional Gaussian noise, alongside a novel mean-preserving degradation operator to significantly reduce sampling steps and inference time. Additionally, a stepwise mapping and error optimization module (SMEO) was designed to precisely calibrate stepwise features using contextual information, thereby minimizing cumulative errors during reconstruction. A retrospective dataset of 50 MP-SPECT scans from 36 patients was used, each gated into 8 (CG-8) or 16 (CG-16) cardiac phases, generating 400/800 image pairs for CG-8/CG-16, respectively. The dataset was divided into training (35 scans), validation (5 scans), and testing (10 scans). Peak signal-to-noise ratio (PSNR), structural similarity (SSIM), normalized mean square error (NMSE), joint histogram, linear regression analysis and a paired two-tail t-test were employed for quantitative evaluation. Two board-certified nuclear medicine physicians performed a blinded and randomized reader study on resulted images. Images were rated on 5-point Likert scales for image quality and diagnostic confidence, with significance evaluated by Wilcoxon signed-rank tests.

The SMGDiff model with 5 diffusion steps (SMGDiff-5) achieved the best overall performance across all evaluation metrics for both gating configurations. SMGDiff-5 also demonstrated superior computational efficiency, requiring only 0.024 s per slice compared to 4.982 s for the 1000-step diffusion model. Furthermore, SMGDiff-5 significantly outperformed established deep learning methods including CNN, U-Net, GAN, and the denoising diffusion probabilistic model, as evidenced by higher PSNR and SSIM and lower NMSE (p < 0.05). Joint histogram and linear regression analyses confirmed these quantitative findings. Reader study results aligned with the quantitative trends which SMGDiff-5 received the highest or near-reference ratings for image quality (CG-8 4.725; CG-16 4.550) and diagnostic confidence (CG-8 4.600; CG-16 4.525), clearly above original gated images and comparable to static MP-SPECT.

The proposed SMGDiff-5 model provides robust and efficient denoising of CG MP-SPECT images, offering superior performance compared to traditional deep learning methods with significantly reduced computational demand.